"""CS Wave 3：客户 Agent LangGraph 编排测试（mock LLM/RAG/工具/Redis）。""" from __future__ import annotations import pytest from app.model.schemas import AuthContext from app.service import customer_service as cs from app.service.profile_service import CustomerMemoryService # --------------------------------------------------------------------------- # 测试替身 # --------------------------------------------------------------------------- class FakeMsg: def __init__(self, content: str) -> None: self.content = content class FakeLLM: def __init__(self, responses: list[str]) -> None: self.responses = list(responses) self.calls: list[list[dict]] = [] def invoke(self, messages): self.calls.append(messages) return FakeMsg(self.responses.pop(0) if self.responses else "{}") class FakeRag: def __init__(self, context: str, sources: list[dict] | None = None) -> None: self._context = context self._sources = sources or [] def retrieve(self, intent, query, top_k=5): return self._context, self._sources def _ctx() -> AuthContext: return AuthContext( sub="CUST-9527", token_type="customer", roles=["customer"], trace_id="t1", agent_type="customer", jti="j1", ) @pytest.fixture def env(monkeypatch, fake_redis): """统一替身环境：LLM / 记忆 / 热缓存 / 后台线程同步化 / 归档。""" llm = FakeLLM([]) recorded = {"intents": [], "extract": [], "archive": [], "idle": []} monkeypatch.setattr(cs, "_build_llm", lambda: llm) monkeypatch.setattr(cs, "CustomerMemoryService", lambda: CustomerMemoryService(redis_client=fake_redis)) monkeypatch.setattr(cs, "ProfileHotCache", lambda repo=None: _FakeHotCache()) monkeypatch.setattr(cs, "record_intent", lambda cid, intent: recorded["intents"].append((cid, intent))) monkeypatch.setattr(cs, "_spawn", lambda fn, *a: fn(*a)) # 同步执行，便于断言 monkeypatch.setattr( cs, "extract_profile", lambda cid, window, tid="": recorded["extract"].append((cid, window, tid)) or [], ) monkeypatch.setattr( cs, "archive_session", lambda sid, cid, tid, reason: recorded["archive"].append((sid, cid, tid, reason)) or True, ) monkeypatch.setattr( cs, "archive_idle_sessions", lambda tid, exclude_sid="", limit=None: recorded["idle"].append((tid, exclude_sid)) or [], ) monkeypatch.setattr("app.config.database.get_redis_client", lambda: fake_redis) return {"llm": llm, "fake": fake_redis, "recorded": recorded} class _FakeHotCache: def get_style_tags(self, cid): return {"basic": {"city": {"value": "上海", "source": "user_declared", "confidence": 0.9}}} def get_style_tags_lazy(self, cid): return self.get_style_tags(cid) def invalidate(self, cid): pass # --------------------------------------------------------------------------- # 意图路由（关键词快路由，零 LLM） # --------------------------------------------------------------------------- def test_route_reject_advice_keyword(env): reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "帮我推荐个基金", "s1", "CUST-9527") assert intent == "reject" assert "投资建议" in reply assert transfer is False assert env["llm"].calls == [] # 未消耗 LLM def test_route_transfer_human(env): reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "转人工", "s1", "CUST-9527") assert intent == "transfer_human" assert transfer is True assert "人工客服" in reply def test_route_fallback_on_invalid_llm_label(env): env["llm"].responses = ["what_is_this"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "今天天气不错", "s1", "CUST-9527") assert intent == "fallback" assert reply == cs.FALLBACK_TEXT # --------------------------------------------------------------------------- # 数据查询分支：param_extract → tool_call → interpret # --------------------------------------------------------------------------- def test_holding_query_end_to_end(env, monkeypatch): captured: dict = {} def fake_holdings(cid, repo=None): captured["cid"] = cid return {"tool": "holding_query", "ok": True, "facts": [], "error": None, "fact_text": "您当前持有 3 只产品，合计市值 149,940.00 元。"} monkeypatch.setitem(cs._TOOL_BY_INTENT, "holding_query", fake_holdings) env["llm"].responses = ["您当前持有 3 只产品，合计市值 149,940.00 元，如需了解单只产品可继续询问。"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "我的持仓怎么样", "s1", "CUST-9527") assert intent == "holding_query" assert captured["cid"] == "CUST-9527" assert "149,940.00" in reply assert disc is False and transfer is False def test_transaction_query_month_extraction(env, monkeypatch): captured: dict = {} def fake_trades(cid, months=None, repo=None): captured["months"] = months return {"tool": "transaction_query", "ok": True, "facts": [], "error": None, "fact_text": "近 6 个月共 2 笔交易。"} monkeypatch.setitem(cs._TOOL_BY_INTENT, "transaction_query", fake_trades) env["llm"].responses = [ '{"months": 6, "product_keyword": null, "risk_level": null}', "您近 6 个月共 2 笔交易，均为申购。", ] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "最近半年的流水", "s1", "CUST-9527") assert intent == "transaction_query" assert captured["months"] == 6 assert "2 笔交易" in reply def test_suitability_risk_level_normalized(env, monkeypatch): captured: dict = {} def fake_suit(cid, product_keyword=None, risk_level=None, repo=None): captured["risk_level"] = risk_level captured["product_keyword"] = product_keyword return {"tool": "suitability_check", "ok": True, "facts": [], "error": None, "fact_text": "R3 汇总：匹配 2 只，需双录 5 只，不匹配 1 只。"} monkeypatch.setitem(cs._TOOL_BY_INTENT, "suitability_check", fake_suit) env["llm"].responses = [ '{"months": null, "product_keyword": null, "risk_level": "r3"}', "R3 产品中匹配 2 只、需双录 5 只、不匹配 1 只。", ] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "我能买R3产品吗", "s1", "CUST-9527") assert intent == "suitability_check" assert captured["risk_level"] == "R3" assert captured["product_keyword"] is None def test_tool_error_returns_data_error_text(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "holding_query", lambda cid, repo=None: {"tool": "holding_query", "ok": False, "facts": None, "fact_text": "", "error": "db_down"}, ) reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "我的持仓怎么样", "s1", "CUST-9527") assert intent == "holding_query" assert reply == cs.DATA_ERROR_TEXT def test_interpret_sanitized_on_forbidden_reply(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "holding_query", lambda cid, repo=None: {"tool": "holding_query", "ok": True, "facts": [], "error": None, "fact_text": "您当前持有 3 只产品。"}, ) env["llm"].responses = ["我建议您买入更多高风险产品。"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "我的持仓怎么样", "s1", "CUST-9527") assert transfer is False assert "3 只产品" in reply assert "无法提供投资建议" not in reply def test_interpret_falls_back_to_fact_text_when_llm_down(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "holding_query", lambda cid, repo=None: {"tool": "holding_query", "ok": True, "facts": [], "error": None, "fact_text": "您当前持有 3 只产品。"}, ) def _boom(*a, **kw): raise RuntimeError("llm down") monkeypatch.setattr(cs, "_build_llm", _boom) reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "我的持仓怎么样", "s1", "CUST-9527") assert "3 只产品" in reply # 数值不经 LLM 也能返回 # --------------------------------------------------------------------------- # RAG 分支 # --------------------------------------------------------------------------- def test_rag_generate_with_disclaimer(env, monkeypatch): monkeypatch.setattr(cs, "VisitorRagService", lambda: FakeRag("债券基金以债券为主要投资标的。")) env["llm"].responses = ["product_consult", "债券基金以债券为主要投资标的，风险相对较低。"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "债券基金是什么", "s1", "CUST-9527") assert intent == "product_consult" assert disc is True assert "风险提示" in reply # 画像注入 prompt（红线：仅语气参考） assert "所在城市：上海" in env["llm"].calls[1][1]["content"] def test_rag_empty_falls_back(env, monkeypatch): monkeypatch.setattr(cs, "VisitorRagService", lambda: FakeRag("")) env["llm"].responses = ["faq"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "怎么修改手机号", "s1", "CUST-9527") assert intent == "fallback" assert reply == cs.FALLBACK_TEXT # --------------------------------------------------------------------------- # 闲聊分支（画像语境注入） # --------------------------------------------------------------------------- def test_greeting_chitchat_without_llm(env, monkeypatch): """「你好」走关键词 chit_chat；DeepSeek 不可用时仍返回闲聊降级话术。""" monkeypatch.setattr( cs, "_invoke", lambda _s, _u: (_ for _ in ()).throw(RuntimeError("no deepseek")), ) reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "你好", "s1", "CUST-9527") assert intent == "chit_chat" assert "暂时无法回答" not in reply assert "您好" in reply or "财富助手" in reply assert transfer is False def test_chitchat_with_profile_context(env): env["llm"].responses = ["您好呀，很高兴为您服务。"] reply, disc, intent, transfer = cs.run_customer_chat(_ctx(), "你好呀", "s1", "CUST-9527") assert intent == "chit_chat" assert "您好" in reply assert disc is False user_prompt = env["llm"].calls[0][1]["content"] assert "所在城市：上海" in user_prompt # --------------------------------------------------------------------------- # C-05 / C-11 数据查询扩展 # --------------------------------------------------------------------------- def test_nav_query_keyword(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "nav_query", lambda cid, product_keyword=None, repo=None: { "tool": "nav_query", "ok": True, "facts": [], "fact_text": "单位净值 1.2345（净值日期 2026-09-09）", }, ) env["llm"].responses = ["nav_query", "单位净值 1.2345（净值日期 2026-09-09）"] reply, disc, intent, transfer = cs.run_customer_chat( _ctx(), "005827的净值是多少", "s1", "CUST-9527", ) assert intent == "nav_query" assert transfer is False assert "1.2345" in reply def test_eligible_products_routes_suitability_not_reject(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "suitability_check", lambda cid, product_keyword=None, risk_level=None, repo=None: { "tool": "suitability_check", "ok": True, "facts": [], "fact_text": "按您当前风评，在售产品适当性匹配汇总：", }, ) env["llm"].responses = [ "suitability_check", "按您当前风评，在售产品适当性匹配汇总：", ] reply, disc, intent, transfer = cs.run_customer_chat( _ctx(), "我能买什么产品", "s1", "CUST-9527", ) assert intent == "suitability_check" assert transfer is False # --------------------------------------------------------------------------- # 记忆保存 / 画像抽槽节流 / 归档 # --------------------------------------------------------------------------- def test_save_memory_writes_consult_line(env, monkeypatch): monkeypatch.setitem( cs._TOOL_BY_INTENT, "holding_query", lambda cid, repo=None: {"tool": "holding_query", "ok": True, "facts": [], "error": None, "fact_text": "您持有 1 只产品。"}, ) env["llm"].responses = ["您持有 1 只产品。"] reply, _, intent, _ = cs.run_customer_chat(_ctx(), "我的持仓", "s1", "CUST-9527") mem = CustomerMemoryService(redis_client=env["fake"]) consult = mem.recall("s1", "consult") assert [m["content"] for m in consult] == ["我的持仓", "您持有 1 只产品。"] assert env["recorded"]["intents"] == [("CUST-9527", "holding_query")] def test_profile_extract_triggered_every_5_rounds(env): env["fake"].strings["customer:s1:rounds"] = "4" # 已累计 4 轮 env["llm"].responses = ["您好。"] cs.run_customer_chat(_ctx(), "你好", "s1", "CUST-9527") # 轮数达到 5 → 触发抽槽（env 中已同步化并记录调用），计数清零 assert env["fake"].strings.get("customer:s1:rounds") is None assert len(env["recorded"]["extract"]) == 1 cid, window, tid = env["recorded"]["extract"][0] assert cid == "CUST-9527" assert tid == "t1" assert "你好" in window # 抽槽窗口含本轮对话（双线合并） def test_profile_extract_not_triggered_before_threshold(env): env["fake"].strings["customer:s1:rounds"] = "1" env["llm"].responses = ["您好。"] cs.run_customer_chat(_ctx(), "你好", "s1", "CUST-9527") assert env["fake"].strings.get("customer:s1:rounds") == "2" # 仅计数 assert env["recorded"]["extract"] == [] def test_archive_on_end_session(env): reply, disc, intent, transfer = cs.run_customer_chat( _ctx(), "转人工", "s1", "CUST-9527", end_session=True ) assert intent == "transfer_human" assert env["recorded"]["archive"] == [("s1", "CUST-9527", "t1", "explicit")] # 懒扫描始终执行，排除当前会话 assert env["recorded"]["idle"] == [("t1", "s1")] def test_archive_idle_scan_every_round(env): cs.run_customer_chat(_ctx(), "转人工", "s1", "CUST-9527") assert env["recorded"]["idle"] == [("t1", "s1")] assert env["recorded"]["archive"] == [] # 未显式结束不归档当前会话 # --------------------------------------------------------------------------- # /api/chat 分流（X-Agent-Type: customer） # --------------------------------------------------------------------------- def test_chat_endpoint_customer_branch(client, monkeypatch): captured: dict = {} def fake_run(ctx, message, session_id, customer_id, end_session=False): captured["customer_id"] = customer_id captured["end_session"] = end_session return ("您好，我是您的客服助手。", False, "chit_chat", False) monkeypatch.setattr("app.api.chat.run_customer_chat", fake_run) login = client.post("/api/auth/login", json={"actor_id": "CUST-9527", "token_type": "customer"}) token = login.json()["data"]["access_token"] resp = client.post( "/api/chat", json={"message": "你好", "end_session": True}, headers={"Authorization": f"Bearer {token}", "X-Agent-Type": "customer"}, ) assert resp.status_code == 200 body = resp.json() assert body["agent_type"] == "customer" assert body["intent"] == "chit_chat" assert body["transfer_to_human"] is False assert captured["customer_id"] == "CUST-9527" assert captured["end_session"] is True def test_chat_endpoint_non_customer_unchanged(client, monkeypatch): """非 customer 分流仍走 agent_service；仅断言响应不含客服专属字段。""" monkeypatch.setattr( "app.api.chat.agent_service.chat", lambda *a, **kw: {"reply": "ok", "has_disclaimer": False}, ) login = client.post("/api/auth/login", json={"actor_id": "STAFF-20001", "token_type": "staff"}) token = login.json()["data"]["access_token"] resp = client.post( "/api/chat", json={"message": "hello"}, headers={"Authorization": f"Bearer {token}", "X-Agent-Type": "analyst"}, ) assert resp.status_code == 200 body = resp.json() assert body.get("intent") is None assert body.get("transfer_to_human") is False